[AArch64] Fix vsqadd scalar intrinsics operands

[dnsampaio]

Summary

Change the vsqadd scalar instrinsics to have the second argument as signed values, not unsigned,
accordingly to https://developer.arm.com/architectures/instruction-sets/simd-isas/neon/intrinsics

The existing unsigned argument can cause faulty code as negative float to unsigned conversion is
undefined, which llvm/clang optimizes away.

[john.brawn]

LGTM, with a couple of nitpicks.

The existing unsigned argument can cause faulty code as float to unsigned conversion is undefined,
which llvm/clang optimizes away.

It's specifically negative float to unsigned that's undefined.

test/CodeGen/aarch64-neon-vsqadd-float-conversion.c

2 ↗

(On Diff #208137)

I don't think -fallow-half-arguments-and-returns is necessary.

cfe/trunk/include/clang/Basic/arm_neon.td

 def SCALAR_SQNEG : SInst<"vqneg", "ss", "ScSsSiSl">;
 
 ////////////////////////////////////////////////////////////////////////////////
 // Scalar Signed Saturating Accumulated of Unsigned Value
 def SCALAR_SUQADD : SInst<"vuqadd", "sss", "ScSsSiSl">;
 
 ////////////////////////////////////////////////////////////////////////////////
 // Scalar Unsigned Saturating Accumulated of Signed Value
-def SCALAR_USQADD : SInst<"vsqadd", "sss", "SUcSUsSUiSUl">;
+def SCALAR_USQADD : SInst<"vsqadd", "ss$", "SUcSUsSUiSUl">;
 
 ////////////////////////////////////////////////////////////////////////////////
 // Signed Saturating Doubling Multiply-Add Long
 def SCALAR_SQDMLAL : SInst<"vqdmlal", "rrss", "SsSi">;
 
 ////////////////////////////////////////////////////////////////////////////////
 // Signed Saturating Doubling Multiply-Subtract Long
 def SCALAR_SQDMLSL : SInst<"vqdmlsl", "rrss", "SsSi">;

cfe/trunk/test/CodeGen/aarch64-neon-intrinsics.c

 }
 
 // CHECK-LABEL: @test_vsqaddb_u8(
 // CHECK:   [[TMP0:%.*]] = insertelement <8 x i8> undef, i8 %a, i64 0
 // CHECK:   [[TMP1:%.*]] = insertelement <8 x i8> undef, i8 %b, i64 0
 // CHECK:   [[VSQADDB_U8_I:%.*]] = call <8 x i8> @llvm.aarch64.neon.usqadd.v8i8(<8 x i8> [[TMP0]], <8 x i8> [[TMP1]])
 // CHECK:   [[TMP2:%.*]] = extractelement <8 x i8> [[VSQADDB_U8_I]], i64 0
 // CHECK:   ret i8 [[TMP2]]
-uint8_t test_vsqaddb_u8(uint8_t a, uint8_t b) {
+uint8_t test_vsqaddb_u8(uint8_t a, int8_t b) {
   return (uint8_t)vsqaddb_u8(a, b);
 }
 
 // CHECK-LABEL: @test_vsqaddh_u16(
 // CHECK:   [[TMP0:%.*]] = insertelement <4 x i16> undef, i16 %a, i64 0
 // CHECK:   [[TMP1:%.*]] = insertelement <4 x i16> undef, i16 %b, i64 0
 // CHECK:   [[VSQADDH_U16_I:%.*]] = call <4 x i16> @llvm.aarch64.neon.usqadd.v4i16(<4 x i16> [[TMP0]], <4 x i16> [[TMP1]])
 // CHECK:   [[TMP2:%.*]] = extractelement <4 x i16> [[VSQADDH_U16_I]], i64 0
 // CHECK:   ret i16 [[TMP2]]
-uint16_t test_vsqaddh_u16(uint16_t a, uint16_t b) {
+uint16_t test_vsqaddh_u16(uint16_t a, int16_t b) {
   return (uint16_t)vsqaddh_u16(a, b);
 }
 
 // CHECK-LABEL: @test_vsqadds_u32(
 // CHECK:   [[VSQADDS_U32_I:%.*]] = call i32 @llvm.aarch64.neon.usqadd.i32(i32 %a, i32 %b)
 // CHECK:   ret i32 [[VSQADDS_U32_I]]
-uint32_t test_vsqadds_u32(uint32_t a, uint32_t b) {
+uint32_t test_vsqadds_u32(uint32_t a, int32_t b) {
   return (uint32_t)vsqadds_u32(a, b);
 }
 
 // CHECK-LABEL: @test_vsqaddd_u64(
 // CHECK:   [[VSQADDD_U64_I:%.*]] = call i64 @llvm.aarch64.neon.usqadd.i64(i64 %a, i64 %b)
 // CHECK:   ret i64 [[VSQADDD_U64_I]]
-uint64_t test_vsqaddd_u64(uint64_t a, uint64_t b) {
+uint64_t test_vsqaddd_u64(uint64_t a, int64_t b) {
   return (uint64_t)vsqaddd_u64(a, b);
 }
 
 // CHECK-LABEL: @test_vqdmlalh_s16(
 // CHECK:   [[TMP0:%.*]] = insertelement <4 x i16> undef, i16 %b, i64 0
 // CHECK:   [[TMP1:%.*]] = insertelement <4 x i16> undef, i16 %c, i64 0
 // CHECK:   [[VQDMLXL_I:%.*]] = call <4 x i32> @llvm.aarch64.neon.sqdmull.v4i32(<4 x i16> [[TMP0]], <4 x i16> [[TMP1]])
 // CHECK:   [[LANE0_I:%.*]] = extractelement <4 x i32> [[VQDMLXL_I]], i64 0

cfe/trunk/test/CodeGen/aarch64-neon-vsqadd-float-conversion.c

+// RUN: %clang_cc1 -triple arm64-none-linux-gnu -target-feature +neon \
+// RUN:  -S -disable-O0-optnone -emit-llvm -o - %s | opt -S -mem2reg -dce \
+// RUN: | FileCheck %s
+
+#include <arm_neon.h>
+
+// Check float conversion is accepted for int argument
+uint8_t test_vsqaddb_u8(){
+  return vsqaddb_u8(1, -1.0f);
+}
+
+uint16_t test_vsqaddh_u16() {
+  return vsqaddh_u16(1, -1.0f);
+}
+
+uint32_t test_vsqadds_u32() {
+  return vsqadds_u32(1, -1.0f);
+}
+
+uint64_t test_vsqaddd_u64() {
+  return vsqaddd_u64(1, -1.0f);
+}
+
+// CHECK-LABEL: @test_vsqaddb_u8()
+// CHECK: entry:
+// CHECK-NEXT: [[T0:%.*]] = insertelement <8 x i8> undef, i8 1, i64 0
+// CHECK-NEXT: [[T1:%.*]] = insertelement <8 x i8> undef, i8 -1, i64 0
+// CHECK-NEXT: [[V:%.*]] = call <8 x i8> @llvm.aarch64.neon.usqadd.v8i8(<8 x i8> [[T0]], <8 x i8> [[T1]])
+// CHECK-NEXT: [[R:%.*]] = extractelement <8 x i8> [[V]], i64 0
+// CHECK-NEXT: ret i8 [[R]]
+
+// CHECK-LABEL: @test_vsqaddh_u16()
+// CHECK: entry:
+// CHECK-NEXT: [[T0:%.*]] = insertelement <4 x i16> undef, i16 1, i64 0
+// CHECK-NEXT: [[T1:%.*]] = insertelement <4 x i16> undef, i16 -1, i64 0
+// CHECK-NEXT: [[V:%.*]]  = call <4 x i16> @llvm.aarch64.neon.usqadd.v4i16(<4 x i16> [[T0]], <4 x i16> [[T1]])
+// CHECK-NEXT: [[R:%.*]] = extractelement <4 x i16> [[V]], i64 0
+// CHECK-NEXT: ret i16 [[R]]
+
+// CHECK-LABEL: @test_vsqadds_u32()
+// CHECK: entry:
+// CHECK-NEXT: [[V:%.*]] = call i32 @llvm.aarch64.neon.usqadd.i32(i32 1, i32 -1)
+// CHECK-NEXT: ret i32 [[V]]
+
+// CHECK-LABEL: @test_vsqaddd_u64()
+// CHECK: entry:
+// CHECK-NEXT: [[V:%.*]] = call i64 @llvm.aarch64.neon.usqadd.i64(i64 1, i64 -1)
+// CHECK-NEXT: ret i64 [[V]]
+